AI safety guardrails stripped fast & Chrome brings AI on-device - Tech News (May 27, 2026)

First up, AI safety—and a reminder that “guardrails” don’t always travel well. The Financial Times, working with the AI safety group Alice, showed that protections on downloadable open-source models from companies like Meta and Google can be stripped quickly. In their demos, modified versions produced content the originals refused, including malware and other clearly harmful material. The bigger takeaway is political, not technical: once a model is out in the wild, safety controls become much harder to enforce at the developer level.

Staying with AI, Google Chrome is making a clear pitch to developers: add AI features directly into web apps without shipping user data to the cloud. At Google I/O, the Chrome team showcased “built-in AI” capabilities that can run on a user’s device, which can cut latency, reduce inference costs, and keep sensitive text local. The practical angle here is mainstream: things like drafting, rewriting, summarizing, and generating structured outputs for moderation or tagging—built into ordinary browser-based workflows.

And Google’s broader AI direction is coming into sharper focus. In a post–I/O interview, CEO Sundar Pichai described reorganizing teams to move faster in an “agentic” era—where Search isn’t just a list of links, but increasingly a place where tasks get done. That’s also why publishers are anxious about a future where AI answers reduce referrals. Google insists links will remain part of the experience, but the tension is clear: AI convenience versus the web’s traffic-based business model.

On AI infrastructure, Nvidia had a strong quarter—and then used the earnings call to telegraph its next land grab: data-center CPUs. CEO Jensen Huang framed CPUs as increasingly important for coordinating agentic AI systems, not just feeding GPUs. If Nvidia can make CPUs a meaningful line of business, it tightens the company’s grip on the entire AI server stack and turns Intel and AMD competition into a much more direct, day-to-day fight.

Another signal that the AI buildout is reshaping old markets: Micron’s valuation briefly crossed the trillion-dollar mark after fresh optimism that large AI customers may lock in longer-term memory supply deals. Memory has historically been a boom-and-bust business. The new bet is that AI demand—especially in big training and inference clusters—could smooth out the cycles and make revenue more predictable than it used to be.

A useful counterpoint arrived from the research world. One observer at the MLSys conference argued the field is currently obsessed with efficiency—faster training, cheaper inference, more specialized chips—and warned that this can narrow experimentation. The concern is that if hardware and software get optimized too aggressively for today’s “frozen model” approach, it could become harder to explore systems that learn continuously after deployment. It’s a timely reminder: infrastructure choices can quietly decide which ideas are feasible.

Now for something more hands-on in AI: Hugging Face released LeRobot Humanoid, an open-source bipedal “legs” platform built around 3D-printed parts and off-the-shelf components. The company is not pitching a sleek demo humanoid—it’s pitching a robot you can understand, fix, instrument, and modify. What makes it especially interesting is the full loop: tools for simulation, calibration, and control, designed so behaviors trained in simulation can be tested on real hardware and then used to improve the next round of training.

In space and markets, SpaceX’s newly filed S-1 is packed with big promises—and a governance structure that will make traditional public-market investors pause. The filing splits the company into space launch, Starlink connectivity, and an AI segment after an early-2026 merger with xAI. Starlink is positioned as the cash engine, while Starship spending keeps the launch side in the red. SpaceX also floated an ambitious “orbital compute” vision—space-based data centers—making Starship’s success the key dependency for multiple future businesses. Separately, the prospectus spells out Musk-friendly voting and pay terms that legal experts say heavily tilt power away from outside shareholders.

SpaceX also picked up another commercial win: American Airlines says it will add Starlink Wi‑Fi to roughly 500 narrow-body Airbus jets starting early next year. Airlines are increasingly treating connectivity as a core feature, not a luxury add-on, and this adds to Starlink’s momentum just as SpaceX gears up for what could be one of the biggest IPOs ever.

NASA, meanwhile, is laying down its own long runway. The agency unveiled plans for three uncrewed lunar landings in 2026 as the start of a major push toward a moon base near the south pole, with Blue Origin tapped for the first cargo landing slot. NASA also outlined future crew mobility rovers and drone-like scouting missions to map terrain and help define operational “perimeters,” a concept that intersects with sensitive questions about safety zones and international norms. And on the computing side, NASA says its next-generation High Performance Spaceflight Computing processor has cleared an early round of environmental testing—important groundwork for more autonomous spacecraft and robots operating far from Earth.

Two quick items on geopolitics and energy. First, Huawei claims a chip-design breakthrough that could help it approach cutting-edge semiconductor density within a few years by leaning harder into 3D structures—though there’s no independent performance proof yet, and heat and toolchain constraints remain major hurdles. Second, the Trump administration is advancing a plan to transfer surplus weapons-grade plutonium to private nuclear startups to be converted into reactor fuel. Supporters call it a clever disposal-and-supply solution; critics warn it could raise serious security and nonproliferation risks.

In tech’s shifting middle class of platforms, Stack Overflow’s traffic story keeps getting more stark. The number of new questions has fallen back to levels reminiscent of its earliest days, as developers turn to AI assistants for quick answers. Yet the business is adapting: more focus on enterprise offerings and licensing its archive—an odd loop where AI tools draw attention away from the community, while still depending on its historical knowledge as valuable training and product input.

And Dropbox is changing leadership. Founder Drew Houston is stepping down as CEO after nearly two decades, moving toward executive chairman while product leader Ashraf Alkarmi transitions into the top job. The backdrop is familiar: intense competition in cloud software, slower growth, and the looming question of how generative AI reshapes subscription products. Dropbox is betting that AI-driven search and content tools can help it stay differentiated—while Houston himself says he plans to pursue a new AI-focused venture outside the company.

Finally, health tech—where the theme is more continuous monitoring and fewer daily interventions. Researchers trialed a wearable ultrasound patch that can image a fetus for hours and track blood flow in real time, which could help spot complications that short, intermittent scans might miss. And Eli Lilly shared early Phase 1 results for a one-time gene-editing treatment that substantially lowered LDL cholesterol at a higher dose, with no treatment-related serious adverse events reported so far. Both stories point to a future where care becomes more proactive, and in some cases, less dependent on constant patient adherence.